/*
Stockfish, a UCI chess playing engine derived from Glaurung 2.1
Copyright (C) 2004-2008 Tord Romstad (Glaurung author)
- Copyright (C) 2008-2010 Marco Costalba, Joona Kiiski, Tord Romstad
+ Copyright (C) 2008-2012 Marco Costalba, Joona Kiiski, Tord Romstad
Stockfish is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
+#include <cassert>
#include <iostream>
+#include "movegen.h"
+#include "search.h"
#include "thread.h"
#include "ucioption.h"
-ThreadsManager ThreadsMgr; // Global object definition
+using namespace Search;
-namespace {
+ThreadsManager Threads; // Global object
- // init_thread() is the function which is called when a new thread is
- // launched. It simply calls the idle_loop() function with the supplied
- // threadID. There are two versions of this function; one for POSIX
- // threads and one for Windows threads.
+namespace { extern "C" {
-#if !defined(_MSC_VER)
+ // start_routine() is the C function which is called when a new thread
+ // is launched. It is a wrapper to member function pointed by start_fn.
- void* init_thread(void* threadID) {
+ long start_routine(Thread* th) { (th->*(th->start_fn))(); return 0; }
- ThreadsMgr.idle_loop(*(int*)threadID, NULL);
- return NULL;
- }
-
-#else
+} }
- DWORD WINAPI init_thread(LPVOID threadID) {
- ThreadsMgr.idle_loop(*(int*)threadID, NULL);
- return 0;
- }
+// Thread c'tor starts a newly-created thread of execution that will call
+// the idle loop function pointed by start_fn going immediately to sleep.
-#endif
+Thread::Thread(Fn fn) {
-}
+ is_searching = do_exit = false;
+ maxPly = splitPointsCnt = 0;
+ curSplitPoint = NULL;
+ start_fn = fn;
+ threadID = Threads.size();
+ do_sleep = (fn != &Thread::main_loop); // Avoid a race with start_searching()
-// read_uci_options() updates number of active threads and other internal
-// parameters according to the UCI options values. It is called before
-// to start a new search.
+ lock_init(sleepLock);
+ cond_init(sleepCond);
-void ThreadsManager::read_uci_options() {
+ for (int j = 0; j < MAX_SPLITPOINTS_PER_THREAD; j++)
+ lock_init(splitPoints[j].lock);
- maxThreadsPerSplitPoint = Options["Maximum Number of Threads per Split Point"].value<int>();
- minimumSplitDepth = Options["Minimum Split Depth"].value<int>() * ONE_PLY;
- useSleepingThreads = Options["Use Sleeping Threads"].value<bool>();
- activeThreads = Options["Threads"].value<int>();
+ if (!thread_create(handle, start_routine, this))
+ {
+ std::cerr << "Failed to create thread number " << threadID << std::endl;
+ ::exit(EXIT_FAILURE);
+ }
}
-// init_threads() is called during startup. Initializes locks and condition
-// variables and launches all threads sending them immediately to sleep.
+// Thread d'tor waits for thread termination before to return.
-void ThreadsManager::init_threads() {
+Thread::~Thread() {
- int arg[MAX_THREADS];
+ assert(do_sleep);
- // This flag is needed to properly end the threads when program exits
- allThreadsShouldExit = false;
+ do_exit = true; // Search must be already finished
+ wake_up();
- // Threads will sent to sleep as soon as created, only main thread is kept alive
- activeThreads = 1;
- threads[0].state = THREAD_SEARCHING;
+ thread_join(handle); // Wait for thread termination
- // Allocate pawn and material hash tables for main thread
- init_hash_tables();
+ lock_destroy(sleepLock);
+ cond_destroy(sleepCond);
- lock_init(&mpLock);
+ for (int j = 0; j < MAX_SPLITPOINTS_PER_THREAD; j++)
+ lock_destroy(splitPoints[j].lock);
+}
- // Initialize thread and split point locks
- for (int i = 0; i < MAX_THREADS; i++)
- {
- lock_init(&threads[i].sleepLock);
- cond_init(&threads[i].sleepCond);
- for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
- lock_init(&(threads[i].splitPoints[j].lock));
+// Thread::timer_loop() is where the timer thread waits maxPly milliseconds and
+// then calls check_time(). If maxPly is 0 thread sleeps until is woken up.
+extern void check_time();
+
+void Thread::timer_loop() {
+
+ while (!do_exit)
+ {
+ lock_grab(sleepLock);
+ timed_wait(sleepCond, sleepLock, maxPly ? maxPly : INT_MAX);
+ lock_release(sleepLock);
+ check_time();
}
+}
+
- // Create and startup all the threads but the main that is already running
- for (int i = 1; i < MAX_THREADS; i++)
+// Thread::main_loop() is where the main thread is parked waiting to be started
+// when there is a new search. Main thread will launch all the slave threads.
+
+void Thread::main_loop() {
+
+ while (true)
{
- threads[i].state = THREAD_INITIALIZING;
- arg[i] = i;
-
-#if !defined(_MSC_VER)
- pthread_t pthread[1];
- bool ok = (pthread_create(pthread, NULL, init_thread, (void*)(&arg[i])) == 0);
- pthread_detach(pthread[0]);
-#else
- bool ok = (CreateThread(NULL, 0, init_thread, (LPVOID)(&arg[i]), 0, NULL) != NULL);
-#endif
- if (!ok)
+ lock_grab(sleepLock);
+
+ do_sleep = true; // Always return to sleep after a search
+ is_searching = false;
+
+ while (do_sleep && !do_exit)
{
- std::cout << "Failed to create thread number " << i << std::endl;
- exit(EXIT_FAILURE);
+ cond_signal(Threads.sleepCond); // Wake up UI thread if needed
+ cond_wait(sleepCond, sleepLock);
}
- // Wait until the thread has finished launching and is gone to sleep
- while (threads[i].state == THREAD_INITIALIZING) {}
+ lock_release(sleepLock);
+
+ if (do_exit)
+ return;
+
+ is_searching = true;
+
+ Search::think();
}
}
-// exit_threads() is called when the program exits. It makes all the
-// helper threads exit cleanly.
+// Thread::wake_up() wakes up the thread, normally at the beginning of the search
+// or, if "sleeping threads" is used at split time.
-void ThreadsManager::exit_threads() {
+void Thread::wake_up() {
- // Force the woken up threads to exit idle_loop() and hence terminate
- allThreadsShouldExit = true;
+ lock_grab(sleepLock);
+ do_sleep = false;
+ cond_signal(sleepCond);
+ lock_release(sleepLock);
+}
- for (int i = 0; i < MAX_THREADS; i++)
- {
- // Wake up all the threads and waits for termination
- if (i != 0)
- {
- threads[i].wake_up();
- while (threads[i].state != THREAD_TERMINATED) {}
- }
- // Now we can safely destroy the locks and wait conditions
- lock_destroy(&threads[i].sleepLock);
- cond_destroy(&threads[i].sleepCond);
+// Thread::wait_for_stop_or_ponderhit() is called when the maximum depth is
+// reached while the program is pondering. The point is to work around a wrinkle
+// in the UCI protocol: When pondering, the engine is not allowed to give a
+// "bestmove" before the GUI sends it a "stop" or "ponderhit" command. We simply
+// wait here until one of these commands (that raise StopRequest) is sent and
+// then return, after which the bestmove and pondermove will be printed.
- for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
- lock_destroy(&(threads[i].splitPoints[j].lock));
- }
+void Thread::wait_for_stop_or_ponderhit() {
+
+ Signals.stopOnPonderhit = true;
- lock_destroy(&mpLock);
+ lock_grab(sleepLock);
+ while (!Signals.stop) cond_wait(sleepCond, sleepLock);
+ lock_release(sleepLock);
}
-// init_hash_tables() dynamically allocates pawn and material hash tables
-// according to the number of active threads. This avoids preallocating
-// memory for all possible threads if only few are used as, for instance,
-// on mobile devices where memory is scarce and allocating for MAX_THREADS
-// threads could even result in a crash.
+// Thread::cutoff_occurred() checks whether a beta cutoff has occurred in the
+// current active split point, or in some ancestor of the split point.
-void ThreadsManager::init_hash_tables() {
+bool Thread::cutoff_occurred() const {
- for (int i = 0; i < activeThreads; i++)
- {
- threads[i].pawnTable.init();
- threads[i].materialTable.init();
- }
+ for (SplitPoint* sp = curSplitPoint; sp; sp = sp->parent)
+ if (sp->cutoff)
+ return true;
+
+ return false;
}
-// cutoff_at_splitpoint() checks whether a beta cutoff has occurred in
-// the thread's currently active split point, or in some ancestor of
-// the current split point.
+// Thread::is_available_to() checks whether the thread is available to help the
+// thread with threadID "master" at a split point. An obvious requirement is that
+// thread must be idle. With more than two threads, this is not sufficient: If
+// the thread is the master of some active split point, it is only available as a
+// slave to the threads which are busy searching the split point at the top of
+// "slave"'s split point stack (the "helpful master concept" in YBWC terminology).
-bool ThreadsManager::cutoff_at_splitpoint(int threadID) const {
+bool Thread::is_available_to(int master) const {
- assert(threadID >= 0 && threadID < activeThreads);
+ if (is_searching)
+ return false;
- SplitPoint* sp = threads[threadID].splitPoint;
+ // Make a local copy to be sure doesn't become zero under our feet while
+ // testing next condition and so leading to an out of bound access.
+ int spCnt = splitPointsCnt;
- for ( ; sp && !sp->betaCutoff; sp = sp->parent) {}
- return sp != NULL;
+ // No active split points means that the thread is available as a slave for any
+ // other thread otherwise apply the "helpful master" concept if possible.
+ return !spCnt || (splitPoints[spCnt - 1].slavesMask & (1ULL << master));
}
-// thread_is_available() checks whether the thread with threadID "slave" is
-// available to help the thread with threadID "master" at a split point. An
-// obvious requirement is that "slave" must be idle. With more than two
-// threads, this is not by itself sufficient: If "slave" is the master of
-// some active split point, it is only available as a slave to the other
-// threads which are busy searching the split point at the top of "slave"'s
-// split point stack (the "helpful master concept" in YBWC terminology).
+// init() is called at startup. Initializes lock and condition variable and
+// launches requested threads sending them immediately to sleep. We cannot use
+// a c'tor becuase Threads is a static object and we need a fully initialized
+// engine at this point due to allocation of endgames in Thread c'tor.
-bool ThreadsManager::thread_is_available(int slave, int master) const {
+void ThreadsManager::init() {
- assert(slave >= 0 && slave < activeThreads);
- assert(master >= 0 && master < activeThreads);
- assert(activeThreads > 1);
+ cond_init(sleepCond);
+ lock_init(splitLock);
+ timer = new Thread(&Thread::timer_loop);
+ threads.push_back(new Thread(&Thread::main_loop));
+ read_uci_options();
+}
- if (threads[slave].state != THREAD_AVAILABLE || slave == master)
- return false;
- // Make a local copy to be sure doesn't change under our feet
- int localActiveSplitPoints = threads[slave].activeSplitPoints;
+// d'tor cleanly terminates the threads when the program exits.
- // No active split points means that the thread is available as
- // a slave for any other thread.
- if (localActiveSplitPoints == 0 || activeThreads == 2)
- return true;
+ThreadsManager::~ThreadsManager() {
- // Apply the "helpful master" concept if possible. Use localActiveSplitPoints
- // that is known to be > 0, instead of threads[slave].activeSplitPoints that
- // could have been set to 0 by another thread leading to an out of bound access.
- if (threads[slave].splitPoints[localActiveSplitPoints - 1].slaves[master])
- return true;
+ for (int i = 0; i < size(); i++)
+ delete threads[i];
- return false;
+ delete timer;
+ lock_destroy(splitLock);
+ cond_destroy(sleepCond);
+}
+
+
+// read_uci_options() updates internal threads parameters from the corresponding
+// UCI options and creates/destroys threads to match the requested number. Thread
+// objects are dynamically allocated to avoid creating in advance all possible
+// threads, with included pawns and material tables, if only few are used.
+
+void ThreadsManager::read_uci_options() {
+
+ maxThreadsPerSplitPoint = Options["Max Threads per Split Point"];
+ minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY;
+ useSleepingThreads = Options["Use Sleeping Threads"];
+ int requested = Options["Threads"];
+
+ assert(requested > 0);
+
+ while (size() < requested)
+ threads.push_back(new Thread(&Thread::idle_loop));
+
+ while (size() > requested)
+ {
+ delete threads.back();
+ threads.pop_back();
+ }
+}
+
+
+// wake_up() is called before a new search to start the threads that are waiting
+// on the sleep condition and to reset maxPly. When useSleepingThreads is set
+// threads will be woken up at split time.
+
+void ThreadsManager::wake_up() const {
+
+ for (int i = 0; i < size(); i++)
+ {
+ threads[i]->maxPly = 0;
+
+ if (!useSleepingThreads)
+ threads[i]->wake_up();
+ }
+}
+
+
+// sleep() is called after the search finishes to ask all the threads but the
+// main one to go waiting on a sleep condition.
+
+void ThreadsManager::sleep() const {
+
+ for (int i = 1; i < size(); i++) // Main thread will go to sleep by itself
+ threads[i]->do_sleep = true; // to avoid a race with start_searching()
}
-// available_thread_exists() tries to find an idle thread which is available as
-// a slave for the thread with threadID "master".
+// available_slave_exists() tries to find an idle thread which is available as
+// a slave for the thread with threadID 'master'.
-bool ThreadsManager::available_thread_exists(int master) const {
+bool ThreadsManager::available_slave_exists(int master) const {
- assert(master >= 0 && master < activeThreads);
- assert(activeThreads > 1);
+ assert(master >= 0 && master < size());
- for (int i = 0; i < activeThreads; i++)
- if (thread_is_available(i, master))
+ for (int i = 0; i < size(); i++)
+ if (threads[i]->is_available_to(master))
return true;
return false;
// split() does the actual work of distributing the work at a node between
-// several available threads. If it does not succeed in splitting the
-// node (because no idle threads are available, or because we have no unused
-// split point objects), the function immediately returns. If splitting is
-// possible, a SplitPoint object is initialized with all the data that must be
-// copied to the helper threads and we tell our helper threads that they have
-// been assigned work. This will cause them to instantly leave their idle loops and
-// call search().When all threads have returned from search() then split() returns.
+// several available threads. If it does not succeed in splitting the node
+// (because no idle threads are available, or because we have no unused split
+// point objects), the function immediately returns. If splitting is possible, a
+// SplitPoint object is initialized with all the data that must be copied to the
+// helper threads and then helper threads are told that they have been assigned
+// work. This will cause them to instantly leave their idle loops and call
+// search(). When all threads have returned from search() then split() returns.
template <bool Fake>
-void ThreadsManager::split(Position& pos, SearchStack* ss, Value* alpha, const Value beta,
- Value* bestValue, Depth depth, Move threatMove,
- int moveCount, MovePicker* mp, bool pvNode) {
- assert(pos.is_ok());
- assert(*bestValue >= -VALUE_INFINITE);
- assert(*bestValue <= *alpha);
- assert(*alpha < beta);
+Value ThreadsManager::split(Position& pos, Stack* ss, Value alpha, Value beta,
+ Value bestValue, Move* bestMove, Depth depth,
+ Move threatMove, int moveCount, MovePicker* mp, int nodeType) {
+ assert(pos.pos_is_ok());
+ assert(bestValue > -VALUE_INFINITE);
+ assert(bestValue <= alpha);
+ assert(alpha < beta);
assert(beta <= VALUE_INFINITE);
assert(depth > DEPTH_ZERO);
- assert(pos.thread() >= 0 && pos.thread() < activeThreads);
- assert(activeThreads > 1);
- int i, master = pos.thread();
- Thread& masterThread = threads[master];
+ int master = pos.thread();
+ Thread& masterThread = *threads[master];
+
+ if (masterThread.splitPointsCnt >= MAX_SPLITPOINTS_PER_THREAD)
+ return bestValue;
+
+ // Pick the next available split point from the split point stack
+ SplitPoint* sp = &masterThread.splitPoints[masterThread.splitPointsCnt++];
+
+ sp->parent = masterThread.curSplitPoint;
+ sp->master = master;
+ sp->cutoff = false;
+ sp->slavesMask = 1ULL << master;
+ sp->depth = depth;
+ sp->bestMove = *bestMove;
+ sp->threatMove = threatMove;
+ sp->alpha = alpha;
+ sp->beta = beta;
+ sp->nodeType = nodeType;
+ sp->bestValue = bestValue;
+ sp->mp = mp;
+ sp->moveCount = moveCount;
+ sp->pos = &pos;
+ sp->nodes = 0;
+ sp->ss = ss;
+
+ assert(masterThread.is_searching);
+
+ masterThread.curSplitPoint = sp;
+ int slavesCnt = 0;
+
+ // Try to allocate available threads and ask them to start searching setting
+ // is_searching flag. This must be done under lock protection to avoid concurrent
+ // allocation of the same slave by another master.
+ lock_grab(sp->lock);
+ lock_grab(splitLock);
+
+ for (int i = 0; i < size() && !Fake; ++i)
+ if (threads[i]->is_available_to(master))
+ {
+ sp->slavesMask |= 1ULL << i;
+ threads[i]->curSplitPoint = sp;
+ threads[i]->is_searching = true; // Slave leaves idle_loop()
+
+ if (useSleepingThreads)
+ threads[i]->wake_up();
- lock_grab(&mpLock);
+ if (++slavesCnt + 1 >= maxThreadsPerSplitPoint) // Master is always included
+ break;
+ }
- // If no other thread is available to help us, or if we have too many
- // active split points, don't split.
- if ( !available_thread_exists(master)
- || masterThread.activeSplitPoints >= MAX_ACTIVE_SPLIT_POINTS)
+ lock_release(splitLock);
+ lock_release(sp->lock);
+
+ // Everything is set up. The master thread enters the idle loop, from which
+ // it will instantly launch a search, because its is_searching flag is set.
+ // We pass the split point as a parameter to the idle loop, which means that
+ // the thread will return from the idle loop when all slaves have finished
+ // their work at this split point.
+ if (slavesCnt || Fake)
{
- lock_release(&mpLock);
- return;
+ masterThread.idle_loop(sp);
+
+ // In helpful master concept a master can help only a sub-tree of its split
+ // point, and because here is all finished is not possible master is booked.
+ assert(!masterThread.is_searching);
}
- // Pick the next available split point object from the split point stack
- SplitPoint& splitPoint = masterThread.splitPoints[masterThread.activeSplitPoints++];
-
- // Initialize the split point object
- splitPoint.parent = masterThread.splitPoint;
- splitPoint.master = master;
- splitPoint.betaCutoff = false;
- splitPoint.depth = depth;
- splitPoint.threatMove = threatMove;
- splitPoint.alpha = *alpha;
- splitPoint.beta = beta;
- splitPoint.pvNode = pvNode;
- splitPoint.bestValue = *bestValue;
- splitPoint.mp = mp;
- splitPoint.moveCount = moveCount;
- splitPoint.pos = &pos;
- splitPoint.nodes = 0;
- splitPoint.ss = ss;
- for (i = 0; i < activeThreads; i++)
- splitPoint.slaves[i] = 0;
-
- masterThread.splitPoint = &splitPoint;
-
- // If we are here it means we are not available
- assert(masterThread.state != THREAD_AVAILABLE);
-
- int workersCnt = 1; // At least the master is included
-
- // Allocate available threads setting state to THREAD_BOOKED
- for (i = 0; !Fake && i < activeThreads && workersCnt < maxThreadsPerSplitPoint; i++)
- if (thread_is_available(i, master))
- {
- threads[i].state = THREAD_BOOKED;
- threads[i].splitPoint = &splitPoint;
- splitPoint.slaves[i] = 1;
- workersCnt++;
- }
+ // We have returned from the idle loop, which means that all threads are
+ // finished. Note that setting is_searching and decreasing splitPointsCnt is
+ // done under lock protection to avoid a race with Thread::is_available_to().
+ lock_grab(sp->lock); // To protect sp->nodes
+ lock_grab(splitLock);
- assert(Fake || workersCnt > 1);
+ masterThread.is_searching = true;
+ masterThread.splitPointsCnt--;
+ masterThread.curSplitPoint = sp->parent;
+ pos.set_nodes_searched(pos.nodes_searched() + sp->nodes);
+ *bestMove = sp->bestMove;
- // We can release the lock because slave threads are already booked and master is not available
- lock_release(&mpLock);
+ lock_release(splitLock);
+ lock_release(sp->lock);
- // Tell the threads that they have work to do. This will make them leave
- // their idle loop.
- for (i = 0; i < activeThreads; i++)
- if (i == master || splitPoint.slaves[i])
- {
- assert(i == master || threads[i].state == THREAD_BOOKED);
+ return sp->bestValue;
+}
- threads[i].state = THREAD_WORKISWAITING; // This makes the slave to exit from idle_loop()
+// Explicit template instantiations
+template Value ThreadsManager::split<false>(Position&, Stack*, Value, Value, Value, Move*, Depth, Move, int, MovePicker*, int);
+template Value ThreadsManager::split<true>(Position&, Stack*, Value, Value, Value, Move*, Depth, Move, int, MovePicker*, int);
- if (useSleepingThreads && i != master)
- threads[i].wake_up();
- }
- // Everything is set up. The master thread enters the idle loop, from
- // which it will instantly launch a search, because its state is
- // THREAD_WORKISWAITING. We send the split point as a second parameter to the
- // idle loop, which means that the main thread will return from the idle
- // loop when all threads have finished their work at this split point.
- idle_loop(master, &splitPoint);
+// ThreadsManager::set_timer() is used to set the timer to trigger after msec
+// milliseconds. If msec is 0 then timer is stopped.
+
+void ThreadsManager::set_timer(int msec) {
+
+ lock_grab(timer->sleepLock);
+ timer->maxPly = msec;
+ cond_signal(timer->sleepCond); // Wake up and restart the timer
+ lock_release(timer->sleepLock);
+}
- // We have returned from the idle loop, which means that all threads are
- // finished. Update alpha and bestValue, and return.
- lock_grab(&mpLock);
- *alpha = splitPoint.alpha;
- *bestValue = splitPoint.bestValue;
- masterThread.activeSplitPoints--;
- masterThread.splitPoint = splitPoint.parent;
- pos.set_nodes_searched(pos.nodes_searched() + splitPoint.nodes);
+// ThreadsManager::wait_for_search_finished() waits for main thread to go to
+// sleep, this means search is finished. Then returns.
- lock_release(&mpLock);
+void ThreadsManager::wait_for_search_finished() {
+
+ Thread* main = threads[0];
+ lock_grab(main->sleepLock);
+ while (!main->do_sleep) cond_wait(sleepCond, main->sleepLock);
+ lock_release(main->sleepLock);
}
-// Explicit template instantiations
-template void ThreadsManager::split<0>(Position&, SearchStack*, Value*, const Value, Value*, Depth, Move, int, MovePicker*, bool);
-template void ThreadsManager::split<1>(Position&, SearchStack*, Value*, const Value, Value*, Depth, Move, int, MovePicker*, bool);
+
+// ThreadsManager::start_searching() is used by UI thread to wake up the main
+// thread parked in main_loop() and starting a new search. If async is true
+// then function returns immediately, otherwise caller is blocked waiting for
+// the search to finish.
+
+void ThreadsManager::start_searching(const Position& pos, const LimitsType& limits,
+ const std::set<Move>& searchMoves, bool async) {
+ wait_for_search_finished();
+
+ Signals.stopOnPonderhit = Signals.firstRootMove = false;
+ Signals.stop = Signals.failedLowAtRoot = false;
+
+ RootPosition.copy(pos, 0);
+ Limits = limits;
+ RootMoves.clear();
+
+ for (MoveList<MV_LEGAL> ml(pos); !ml.end(); ++ml)
+ if (searchMoves.empty() || searchMoves.count(ml.move()))
+ RootMoves.push_back(RootMove(ml.move()));
+
+ threads[0]->wake_up(); // Start main thread
+
+ if (!async)
+ wait_for_search_finished();
+}